blob: fc3e7ffd189eed43489cdabda0dd4cc54612e900 [file] [log] [blame]
Joonsoo Kimeefa8642014-12-12 16:55:46 -08001#include <linux/mm.h>
2#include <linux/mmzone.h>
3#include <linux/bootmem.h>
4#include <linux/page_ext.h>
5#include <linux/memory.h>
6#include <linux/vmalloc.h>
7#include <linux/kmemleak.h>
Joonsoo Kim48c96a32014-12-12 16:56:01 -08008#include <linux/page_owner.h>
Vladimir Davydov33c3fc72015-09-09 15:35:45 -07009#include <linux/page_idle.h>
Joonsoo Kimeefa8642014-12-12 16:55:46 -080010
11/*
12 * struct page extension
13 *
14 * This is the feature to manage memory for extended data per page.
15 *
16 * Until now, we must modify struct page itself to store extra data per page.
17 * This requires rebuilding the kernel and it is really time consuming process.
18 * And, sometimes, rebuild is impossible due to third party module dependency.
19 * At last, enlarging struct page could cause un-wanted system behaviour change.
20 *
21 * This feature is intended to overcome above mentioned problems. This feature
22 * allocates memory for extended data per page in certain place rather than
23 * the struct page itself. This memory can be accessed by the accessor
24 * functions provided by this code. During the boot process, it checks whether
25 * allocation of huge chunk of memory is needed or not. If not, it avoids
26 * allocating memory at all. With this advantage, we can include this feature
27 * into the kernel in default and can avoid rebuild and solve related problems.
28 *
29 * To help these things to work well, there are two callbacks for clients. One
30 * is the need callback which is mandatory if user wants to avoid useless
31 * memory allocation at boot-time. The other is optional, init callback, which
32 * is used to do proper initialization after memory is allocated.
33 *
34 * The need callback is used to decide whether extended memory allocation is
35 * needed or not. Sometimes users want to deactivate some features in this
36 * boot and extra memory would be unneccessary. In this case, to avoid
37 * allocating huge chunk of memory, each clients represent their need of
38 * extra memory through the need callback. If one of the need callbacks
39 * returns true, it means that someone needs extra memory so that
40 * page extension core should allocates memory for page extension. If
41 * none of need callbacks return true, memory isn't needed at all in this boot
42 * and page extension core can skip to allocate memory. As result,
43 * none of memory is wasted.
44 *
Joonsoo Kim980ac162016-10-07 16:58:27 -070045 * When need callback returns true, page_ext checks if there is a request for
46 * extra memory through size in struct page_ext_operations. If it is non-zero,
47 * extra space is allocated for each page_ext entry and offset is returned to
48 * user through offset in struct page_ext_operations.
49 *
Joonsoo Kimeefa8642014-12-12 16:55:46 -080050 * The init callback is used to do proper initialization after page extension
51 * is completely initialized. In sparse memory system, extra memory is
52 * allocated some time later than memmap is allocated. In other words, lifetime
53 * of memory for page extension isn't same with memmap for struct page.
54 * Therefore, clients can't store extra data until page extension is
55 * initialized, even if pages are allocated and used freely. This could
56 * cause inadequate state of extra data per page, so, to prevent it, client
57 * can utilize this callback to initialize the state of it correctly.
58 */
59
60static struct page_ext_operations *page_ext_ops[] = {
Joonsoo Kime30825f2014-12-12 16:55:49 -080061 &debug_guardpage_ops,
Joonsoo Kim48c96a32014-12-12 16:56:01 -080062#ifdef CONFIG_PAGE_OWNER
63 &page_owner_ops,
64#endif
Vladimir Davydov33c3fc72015-09-09 15:35:45 -070065#if defined(CONFIG_IDLE_PAGE_TRACKING) && !defined(CONFIG_64BIT)
66 &page_idle_ops,
67#endif
Joonsoo Kimeefa8642014-12-12 16:55:46 -080068};
69
70static unsigned long total_usage;
Joonsoo Kim980ac162016-10-07 16:58:27 -070071static unsigned long extra_mem;
Joonsoo Kimeefa8642014-12-12 16:55:46 -080072
73static bool __init invoke_need_callbacks(void)
74{
75 int i;
76 int entries = ARRAY_SIZE(page_ext_ops);
Joonsoo Kim980ac162016-10-07 16:58:27 -070077 bool need = false;
Joonsoo Kimeefa8642014-12-12 16:55:46 -080078
79 for (i = 0; i < entries; i++) {
Joonsoo Kim980ac162016-10-07 16:58:27 -070080 if (page_ext_ops[i]->need && page_ext_ops[i]->need()) {
81 page_ext_ops[i]->offset = sizeof(struct page_ext) +
82 extra_mem;
83 extra_mem += page_ext_ops[i]->size;
84 need = true;
85 }
Joonsoo Kimeefa8642014-12-12 16:55:46 -080086 }
87
Joonsoo Kim980ac162016-10-07 16:58:27 -070088 return need;
Joonsoo Kimeefa8642014-12-12 16:55:46 -080089}
90
91static void __init invoke_init_callbacks(void)
92{
93 int i;
94 int entries = ARRAY_SIZE(page_ext_ops);
95
96 for (i = 0; i < entries; i++) {
97 if (page_ext_ops[i]->init)
98 page_ext_ops[i]->init();
99 }
100}
101
Joonsoo Kim980ac162016-10-07 16:58:27 -0700102static unsigned long get_entry_size(void)
103{
104 return sizeof(struct page_ext) + extra_mem;
105}
106
107static inline struct page_ext *get_entry(void *base, unsigned long index)
108{
109 return base + get_entry_size() * index;
110}
111
Joonsoo Kimeefa8642014-12-12 16:55:46 -0800112#if !defined(CONFIG_SPARSEMEM)
113
114
115void __meminit pgdat_page_ext_init(struct pglist_data *pgdat)
116{
117 pgdat->node_page_ext = NULL;
118}
119
120struct page_ext *lookup_page_ext(struct page *page)
121{
122 unsigned long pfn = page_to_pfn(page);
Joonsoo Kim0b06bb32016-10-07 16:58:24 -0700123 unsigned long index;
Joonsoo Kimeefa8642014-12-12 16:55:46 -0800124 struct page_ext *base;
125
126 base = NODE_DATA(page_to_nid(page))->node_page_ext;
Laura Abbott1414c7f2016-03-15 14:56:30 -0700127#if defined(CONFIG_DEBUG_VM) || defined(CONFIG_PAGE_POISONING)
Joonsoo Kimeefa8642014-12-12 16:55:46 -0800128 /*
129 * The sanity checks the page allocator does upon freeing a
130 * page can reach here before the page_ext arrays are
131 * allocated when feeding a range of pages to the allocator
132 * for the first time during bootup or memory hotplug.
Laura Abbott1414c7f2016-03-15 14:56:30 -0700133 *
134 * This check is also necessary for ensuring page poisoning
135 * works as expected when enabled
Joonsoo Kimeefa8642014-12-12 16:55:46 -0800136 */
137 if (unlikely(!base))
138 return NULL;
139#endif
Joonsoo Kim0b06bb32016-10-07 16:58:24 -0700140 index = pfn - round_down(node_start_pfn(page_to_nid(page)),
Joonsoo Kimeefa8642014-12-12 16:55:46 -0800141 MAX_ORDER_NR_PAGES);
Joonsoo Kim980ac162016-10-07 16:58:27 -0700142 return get_entry(base, index);
Joonsoo Kimeefa8642014-12-12 16:55:46 -0800143}
144
145static int __init alloc_node_page_ext(int nid)
146{
147 struct page_ext *base;
148 unsigned long table_size;
149 unsigned long nr_pages;
150
151 nr_pages = NODE_DATA(nid)->node_spanned_pages;
152 if (!nr_pages)
153 return 0;
154
155 /*
156 * Need extra space if node range is not aligned with
157 * MAX_ORDER_NR_PAGES. When page allocator's buddy algorithm
158 * checks buddy's status, range could be out of exact node range.
159 */
160 if (!IS_ALIGNED(node_start_pfn(nid), MAX_ORDER_NR_PAGES) ||
161 !IS_ALIGNED(node_end_pfn(nid), MAX_ORDER_NR_PAGES))
162 nr_pages += MAX_ORDER_NR_PAGES;
163
Joonsoo Kim980ac162016-10-07 16:58:27 -0700164 table_size = get_entry_size() * nr_pages;
Joonsoo Kimeefa8642014-12-12 16:55:46 -0800165
166 base = memblock_virt_alloc_try_nid_nopanic(
167 table_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
168 BOOTMEM_ALLOC_ACCESSIBLE, nid);
169 if (!base)
170 return -ENOMEM;
171 NODE_DATA(nid)->node_page_ext = base;
172 total_usage += table_size;
173 return 0;
174}
175
176void __init page_ext_init_flatmem(void)
177{
178
179 int nid, fail;
180
181 if (!invoke_need_callbacks())
182 return;
183
184 for_each_online_node(nid) {
185 fail = alloc_node_page_ext(nid);
186 if (fail)
187 goto fail;
188 }
189 pr_info("allocated %ld bytes of page_ext\n", total_usage);
190 invoke_init_callbacks();
191 return;
192
193fail:
194 pr_crit("allocation of page_ext failed.\n");
195 panic("Out of memory");
196}
197
198#else /* CONFIG_FLAT_NODE_MEM_MAP */
199
200struct page_ext *lookup_page_ext(struct page *page)
201{
202 unsigned long pfn = page_to_pfn(page);
203 struct mem_section *section = __pfn_to_section(pfn);
Laura Abbott1414c7f2016-03-15 14:56:30 -0700204#if defined(CONFIG_DEBUG_VM) || defined(CONFIG_PAGE_POISONING)
Joonsoo Kimeefa8642014-12-12 16:55:46 -0800205 /*
206 * The sanity checks the page allocator does upon freeing a
207 * page can reach here before the page_ext arrays are
208 * allocated when feeding a range of pages to the allocator
209 * for the first time during bootup or memory hotplug.
Laura Abbott1414c7f2016-03-15 14:56:30 -0700210 *
211 * This check is also necessary for ensuring page poisoning
212 * works as expected when enabled
Joonsoo Kimeefa8642014-12-12 16:55:46 -0800213 */
214 if (!section->page_ext)
215 return NULL;
216#endif
Joonsoo Kim980ac162016-10-07 16:58:27 -0700217 return get_entry(section->page_ext, pfn);
Joonsoo Kimeefa8642014-12-12 16:55:46 -0800218}
219
220static void *__meminit alloc_page_ext(size_t size, int nid)
221{
222 gfp_t flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN;
223 void *addr = NULL;
224
225 addr = alloc_pages_exact_nid(nid, size, flags);
226 if (addr) {
227 kmemleak_alloc(addr, size, 1, flags);
228 return addr;
229 }
230
231 if (node_state(nid, N_HIGH_MEMORY))
232 addr = vzalloc_node(size, nid);
233 else
234 addr = vzalloc(size);
235
236 return addr;
237}
238
239static int __meminit init_section_page_ext(unsigned long pfn, int nid)
240{
241 struct mem_section *section;
242 struct page_ext *base;
243 unsigned long table_size;
244
245 section = __pfn_to_section(pfn);
246
247 if (section->page_ext)
248 return 0;
249
Joonsoo Kim980ac162016-10-07 16:58:27 -0700250 table_size = get_entry_size() * PAGES_PER_SECTION;
Joonsoo Kimeefa8642014-12-12 16:55:46 -0800251 base = alloc_page_ext(table_size, nid);
252
253 /*
254 * The value stored in section->page_ext is (base - pfn)
255 * and it does not point to the memory block allocated above,
256 * causing kmemleak false positives.
257 */
258 kmemleak_not_leak(base);
259
260 if (!base) {
261 pr_err("page ext allocation failure\n");
262 return -ENOMEM;
263 }
264
265 /*
266 * The passed "pfn" may not be aligned to SECTION. For the calculation
267 * we need to apply a mask.
268 */
269 pfn &= PAGE_SECTION_MASK;
Joonsoo Kim980ac162016-10-07 16:58:27 -0700270 section->page_ext = (void *)base - get_entry_size() * pfn;
Joonsoo Kimeefa8642014-12-12 16:55:46 -0800271 total_usage += table_size;
272 return 0;
273}
274#ifdef CONFIG_MEMORY_HOTPLUG
275static void free_page_ext(void *addr)
276{
277 if (is_vmalloc_addr(addr)) {
278 vfree(addr);
279 } else {
280 struct page *page = virt_to_page(addr);
281 size_t table_size;
282
Joonsoo Kim980ac162016-10-07 16:58:27 -0700283 table_size = get_entry_size() * PAGES_PER_SECTION;
Joonsoo Kimeefa8642014-12-12 16:55:46 -0800284
285 BUG_ON(PageReserved(page));
286 free_pages_exact(addr, table_size);
287 }
288}
289
290static void __free_page_ext(unsigned long pfn)
291{
292 struct mem_section *ms;
293 struct page_ext *base;
294
295 ms = __pfn_to_section(pfn);
296 if (!ms || !ms->page_ext)
297 return;
Joonsoo Kim980ac162016-10-07 16:58:27 -0700298 base = get_entry(ms->page_ext, pfn);
Joonsoo Kimeefa8642014-12-12 16:55:46 -0800299 free_page_ext(base);
300 ms->page_ext = NULL;
301}
302
303static int __meminit online_page_ext(unsigned long start_pfn,
304 unsigned long nr_pages,
305 int nid)
306{
307 unsigned long start, end, pfn;
308 int fail = 0;
309
310 start = SECTION_ALIGN_DOWN(start_pfn);
311 end = SECTION_ALIGN_UP(start_pfn + nr_pages);
312
313 if (nid == -1) {
314 /*
315 * In this case, "nid" already exists and contains valid memory.
316 * "start_pfn" passed to us is a pfn which is an arg for
317 * online__pages(), and start_pfn should exist.
318 */
319 nid = pfn_to_nid(start_pfn);
320 VM_BUG_ON(!node_state(nid, N_ONLINE));
321 }
322
323 for (pfn = start; !fail && pfn < end; pfn += PAGES_PER_SECTION) {
324 if (!pfn_present(pfn))
325 continue;
326 fail = init_section_page_ext(pfn, nid);
327 }
328 if (!fail)
329 return 0;
330
331 /* rollback */
332 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
333 __free_page_ext(pfn);
334
335 return -ENOMEM;
336}
337
338static int __meminit offline_page_ext(unsigned long start_pfn,
339 unsigned long nr_pages, int nid)
340{
341 unsigned long start, end, pfn;
342
343 start = SECTION_ALIGN_DOWN(start_pfn);
344 end = SECTION_ALIGN_UP(start_pfn + nr_pages);
345
346 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
347 __free_page_ext(pfn);
348 return 0;
349
350}
351
352static int __meminit page_ext_callback(struct notifier_block *self,
353 unsigned long action, void *arg)
354{
355 struct memory_notify *mn = arg;
356 int ret = 0;
357
358 switch (action) {
359 case MEM_GOING_ONLINE:
360 ret = online_page_ext(mn->start_pfn,
361 mn->nr_pages, mn->status_change_nid);
362 break;
363 case MEM_OFFLINE:
364 offline_page_ext(mn->start_pfn,
365 mn->nr_pages, mn->status_change_nid);
366 break;
367 case MEM_CANCEL_ONLINE:
368 offline_page_ext(mn->start_pfn,
369 mn->nr_pages, mn->status_change_nid);
370 break;
371 case MEM_GOING_OFFLINE:
372 break;
373 case MEM_ONLINE:
374 case MEM_CANCEL_OFFLINE:
375 break;
376 }
377
378 return notifier_from_errno(ret);
379}
380
381#endif
382
383void __init page_ext_init(void)
384{
385 unsigned long pfn;
386 int nid;
387
388 if (!invoke_need_callbacks())
389 return;
390
391 for_each_node_state(nid, N_MEMORY) {
392 unsigned long start_pfn, end_pfn;
393
394 start_pfn = node_start_pfn(nid);
395 end_pfn = node_end_pfn(nid);
396 /*
397 * start_pfn and end_pfn may not be aligned to SECTION and the
398 * page->flags of out of node pages are not initialized. So we
399 * scan [start_pfn, the biggest section's pfn < end_pfn) here.
400 */
401 for (pfn = start_pfn; pfn < end_pfn;
402 pfn = ALIGN(pfn + 1, PAGES_PER_SECTION)) {
403
404 if (!pfn_valid(pfn))
405 continue;
406 /*
407 * Nodes's pfns can be overlapping.
408 * We know some arch can have a nodes layout such as
409 * -------------pfn-------------->
410 * N0 | N1 | N2 | N0 | N1 | N2|....
Yang Shife53ca52016-05-27 14:27:30 -0700411 *
412 * Take into account DEFERRED_STRUCT_PAGE_INIT.
Joonsoo Kimeefa8642014-12-12 16:55:46 -0800413 */
Yang Shife53ca52016-05-27 14:27:30 -0700414 if (early_pfn_to_nid(pfn) != nid)
Joonsoo Kimeefa8642014-12-12 16:55:46 -0800415 continue;
416 if (init_section_page_ext(pfn, nid))
417 goto oom;
418 }
419 }
420 hotplug_memory_notifier(page_ext_callback, 0);
421 pr_info("allocated %ld bytes of page_ext\n", total_usage);
422 invoke_init_callbacks();
423 return;
424
425oom:
426 panic("Out of memory");
427}
428
429void __meminit pgdat_page_ext_init(struct pglist_data *pgdat)
430{
431}
432
433#endif